AI war rooms can crack legacy code, GEICO's $1,000 day shows

The claim. GEICO senior staff engineer Ford Prior ran three one-day experiments using AI agents to map and modernize legacy claims technology. With $1,000 in tokens per session, the agents produced architecture graphs, API proposals, and in the salvage domain a fully vetted migration plan signed off by product and tech. The highest-value output was the decision-ready plan, not the code.

Why it matters. Prior says the inner loop, writing and building code, has mostly collapsed at GEICO. The outer loop, every decision and review around that code, now limits delivery. Their engineers spent 75% or more of capacity on patches and minor upgrades, leaving a quarter or less for major work. The war room is an attempt to compress that outer loop by assembling experts, a driver, and AI agents around one problem.

How it ran. The setup combined Anthropic's Claude, a code modernization plugin, a not-yet-open-source fleet manager called FleetKit, MCP servers, and GEICO's own context: standards, past decisions, tools, and approval rules. Agents ran in the cloud and on a driver's laptop, created a Slack channel for the war room, and produced HTML worksheets and design documents for humans to review. The driver's machine was the runtime, which is a scaling constraint.

The results. In the rentals domain, the agent found a way to save 8,500 calls per month by connecting rental data. In damages, it built agent-callable APIs for a larger workflow. In salvage, it produced a migration plan with API specs, design mockups, and full approval, and found a production bug the team could not discuss. The artifacts were good enough that teams kept the architecture diagrams.

The trade-offs. Token costs are real. Prior described Claude/Fable creating cron jobs to check on work, which launched new sessions and burned budget. Prompt caching helped more than expected, and internal benchmarks showed Sonnet 5.5 matched or beat Fable on generated code. None of this is a production system; it is three experiments with a vendor's plugin and heavy custom glue.

The caveat. This is a story about process more than product. The reusable insight is the explicit human approval gate, the reusable context, and the artifact trail. The talk's own language, 'factory' instead of canvas and 'divine intervention' for engineers, should be treated as enthusiasm. Engineering leaders should run a bounded pilot before assuming $1,000 days are repeatable.